Norming and bias in tests as the US guidelines waver

In mid September 2026, SHRM, the US society for HR professionals, described a situation that HR teams on both sides of the Atlantic will recognise. Federal authorities in the United States are shifting their position on disparate impact, the term for a seemingly neutral selection method that screens out some groups at a higher rate than others. At the same time, the rulebook employers have used to document their tests is under review. That puts norming and bias in tests back on the agenda.

According to SHRM, the US Department of Justice issued an opinion in June stating that the EEOC's long-standing interpretation of disparate impact liability is unconstitutional. The opinion was requested by EEOC chair Andrea Lucas and argues that the agency has placed too much weight on statistical disparities and too little on intent. The EEOC has also proposed rescinding the Uniform Guidelines on Employee Selection Procedures, the framework from 1978 that employers have relied on for 50 years to evaluate and validate cognitive and personality tests, work samples, structured interviews and other selection procedures.

SHRM's central point is that none of this changes the law. Disparate impact remains part of Title VII, and several states are moving in the opposite direction. New York and New Jersey have expanded liability, Illinois and Maryland have written it into statute, and California, Colorado, Massachusetts and Minnesota recognise it through legislation or case law. Jocelyn King of VirgilHR puts it plainly. Employers should not mistake reduced federal enforcement for reduced discrimination risk.

Two concepts sit at the heart of this debate, and they are often confused. A norm is the reference point a test score is compared against. A raw score of 27 correct answers means nothing until it is set against how a defined group performed on the same test. That norm group can be broad, such as the working population, or narrow, such as applicants for a specific type of role. It can be national or international, recent or decades old. Each of those choices shifts where a candidate lands on the scale, and therefore who moves forward.

Bias is a different question. A test is biased if it systematically measures differently for different groups even when those groups have the same level of the ability or trait being measured. That is not the same as groups scoring differently. A gap in averages can reflect real differences, a poorly matched norm group, or items that work differently for different groups. Telling those three apart requires documentation.

The Uniform Guidelines have never been law in Denmark or anywhere else in Europe. What they have done is set a professional bar for what a test publisher must be able to show when its instrument is used for hiring. Publishers who documented their tests to that standard did so because employers in their largest market required it. If the framework is withdrawn at home, it is a fair question whether the next edition of a manual will still contain the same evidence.

For European employers, the obligations do not move. Indirect discrimination is prohibited under EU equal treatment rules and national legislation, and the EU AI Act places AI systems used in recruitment in the high-risk category. The responsibility sits with the organisation that uses the tool.

That is also where SHRM lands. Cara Crotty, an attorney at Constangy, notes that when a hiring tool screens out applicants in protected groups at disproportionate rates, plaintiffs will generally go after the employer who used the tool to make the decision, not the vendor or developer. She advises employers to understand how an AI system works, to know the underlying data, to confirm that independent validation has been completed, and to make sure contracts cover legal compliance, documentation and data retention.

Lynn Clements of Berkshire Associates recommends that employers keep their validation studies current, confirm that a test is documented for the specific job, and continue monitoring selection outcomes. In her words, the strongest way to demonstrate business necessity remains the same approach employers have used for decades.

Translated into everyday HR practice, this gives a short list of questions. Which norm group is used, how large is it, when was it collected, and does it resemble the candidates we actually test? Has the publisher checked whether the test functions the same way across gender, age and language background? Is there an analysis of how a given cut-off affects different groups? And can the publisher show that the test predicts something related to the job?

The answers belong in the technical manual, not in a sales deck. The American debate is a reminder that rules can change from one year to the next. The need to explain why a candidate was rejected does not.

Source: https://www.shrm.org/topics-tools/news/eeoc-disparate-impact-shift-hr-uncertainty